dsh-lens-lite

Post-edit diagnostics for DeepSeek Harness.

After a file-mutating tool succeeds, this plugin runs the checkers you configured for that file's extension and attaches their output to the same tool result as model-visible context. The next model request therefore already carries the type errors or lint findings the edit introduced, instead of the model discovering them a few turns later.

dsh-lsp covers semantic navigation (definition, references, hover). It does not surface diagnostics, and nothing in the shipped harness runs a type-checker after an edit. This plugin fills that gap.

Install

dsh plugin --profile web add dsh-lens-lite

Then override the row in your profile's cordis.patch.yml with the checkers your toolchain actually has (the shipped row has none — see Configuration).

How it behaves

• Runs only after a successful call to a watched tool (write, edit, str_replace_editor by default).
• Takes the edited path from the tool's canonical path value — the path the filesystem backend actually resolved — and falls back to the file_path argument for tools that declare no such value.
• Runs every checker claiming that file's extension, concurrently.
• Attaches nothing when every checker exits with a configured clean code and prints nothing.
• Never vetoes and never rewrites a call. It delegates the tools/post-execute waterfall first and only folds context onto whatever decision came back, preserving every downstream context's own source and metadata.
• Skips linting entirely when a downstream listener blocked the call — that blocked call is already the feedback the model must react to.
• Honours the tool call's cancellation signal, and terminates a checker's whole process tree when its own timeout expires.

Findings arrive as a notice-form plugin context, so the transcript shows a collapsed one-line summary rather than a wall of compiler output.

Configuration

Every deployment-varying value is a config field. The plugin ships no built-in commands: which checkers exist is a property of your toolchain, not of this package.

- id: lens-lite
  name: dsh-lens-lite
  config:
    tools: [write, edit, str_replace_editor]
    maxDiagnosticChars: 4000
    checkers:
      - name: tsc
        extensions: ['.ts', '.tsx']
        argv: ['npx', '--no-install', 'tsc', '--noEmit']
        cwd: '.'
        timeoutMs: 60000

      - name: eslint
        extensions: ['.ts', '.tsx', '.js', '.jsx']
        argv: ['npx', '--no-install', 'eslint', '--format', 'compact', '{file}']
        cleanExitCodes: [0]

      - name: ruff
        extensions: ['.py']
        argv: ['ruff', 'check', '--output-format', 'concise', '{file}']

      - name: go-vet
        extensions: ['.go']
        argv: ['go', 'vet', '{dir}']
        streams: [stderr]

Checker fields

Field	Default	Meaning
name	— (required)	Label shown to the model above this checker's output
extensions	— (required)	Lowercase, leading-dot extensions this checker claims
argv	— (required)	Executable + arguments. argv[0] resolves against the subprocess provider's scrubbed PATH. Never shell-interpreted
cwd	.	Working directory, resolved against the harness process directory (your workspace root)
timeoutMs	30000	Wall-clock bound; the process tree is terminated when it expires
graceMs	2000	SIGTERM→SIGKILL grace for that termination
maxOutputBytes	65536	In-memory cap per collected stream; overflow keeps the tail
cleanExitCodes	[0]	Exit codes meaning "no findings"
streams	[stdout, stderr]	Which streams carry findings, in concatenation order

argv entries support three placeholders, substituted per run:

Placeholder	Value
{file}	Absolute path of the edited file
{relFile}	That path relative to cwd
{dir}	Its containing directory

Plugin fields

Field	Default	Meaning
tools	[write, edit, str_replace_editor]	Tool names whose successful results are inspected
checkers	[]	Checkers, all consulted; empty means the plugin does nothing (it says so once at load)
maxDiagnosticChars	4000	Cap on model-visible diagnostic text per result; overflow keeps the head, because the first error is usually the cause of the rest

Cost

Whole-project checkers run on every edit. npx tsc --noEmit on a large repo can add seconds to each write. Prefer single-file invocations ({file}) where the tool supports them, and raise timeoutMs only for the checkers that genuinely need it.

Failure behavior

Load-time misconfiguration fails loud: an empty argv, a checker claiming no extensions or reading no streams, a duplicate checker name, or a non-positive bound throws at plugin load.

Environment failures do not fail the edit — the call already succeeded — and become findings instead:

• An unresolvable executable is reported once as checker unavailable: …, then that checker stays quiet for the rest of the fiber's life.
• A timeout is reported as checker timed out after Nms.
• A spawn-level failure is reported as checker failed to start: ….

Extension point

One listener on tools/post-execute, plus ctx.subprocess for spawning. No core changes, no agent-loop changes.

Development

pnpm install --ignore-workspace
pnpm run typecheck
pnpm test
pnpm run build

The test suite drives the real tools/post-execute waterfall against the real local subprocess provider, using node -e programs as checkers, so it covers actual spawning, exit-code classification, stream selection, and timeouts without requiring any toolchain.

License

MIT

Prior art

The post-edit-feedback idea comes from pi-lens (MIT) in the Pi ecosystem. This is an independent implementation against Harness extension points and shares no code with it; it deliberately covers only the run-checkers-after-edit slice, not pi-lens's AST rules, dependency mapping, or triage system.